Research And Implementation Of Camera Pose Estimation Method In Low-light Scene

Geometric consistency in Augmented Reality(AR)requires mobile devices such as cameras be located in the real world to accurately add virtual information.Camera pose estimation is one of the key technologies in AR because it can obtain accurate and real-time camera six degrees of pose from captured images.At present,the technology has the problem of low accuracy in scenes with dramatic light and viewpoint changes,specially,in low-illumination scenes,the extracted keypoint information is limited due to weak light conditions which resulting in poor camera pose estimation.This paper does the following work to solve this problem:1.A feature fusion method for low-illumination images is presented to solve the problem of poor extraction of keypoints from traditional camera pose estimation methods in lowillumination scenes.This method first preprocesses the low-illumination image with light enhancement,then fuses the features extracted with SIFT which is the classic traditional keypoint extraction method and Super Point which belongs to the deep learning method.Finally,it uses the homography transformation characteristics of the input image pairs to propose the homography transformation loss items to enhance the network constraints.This method can extract robust keypoint information and show good performance in low-illumination scenes.2.An adaptive motion difference feature enhancement network is proposed to improve the accuracy of camera pose estimation in scenes with varying illumination and viewpoint in deep learning.The network extends the single-frame features to continuous frames,combines the motion difference feature between frames with the static feature of the input frame as the interframe correlation feature,then an attention module is applied to adaptively focuses the important portion of the network that is conducive to pose estimation.Finally,relative motion loss is adopted as a further constraint to the network.This method can get more accurate camera pose information and achieve competitive results on the open dataset.This paper studies and implements camera pose estimation methods for low-illumination scenes.The experimental results show that,on the one hand,the proposed algorithm improves the ability of keypoint extraction in low-illumination scenes and lays a good data foundation for camera pose estimation.On the other hand,it improves the performance of camera pose estimation in indoor and outdoor scenes.Compared with traditional and deep learning methods,this method has better accuracy and stability.Therefore,the pose estimation method proposed in this paper has certain theoretical significance and application value.